A Lightweight and Optimized BERT-based Intrusion Detection System for Resource-Constrained Network and Logistics Environments

This paper proposes OB-IDS (Optimized BERT-based Intrusion Detection System), a highly efficient and lightweight intrusion detection model specifically designed for resource-constrained environments such as edge devices, IoT gateways, and embedded network appliances.

Modern logistics and supply chain infrastructures increasingly rely on interconnected cyber-physical systems, including warehouse management systems (WMS), transportation management systems (TMS), automated sorting machines, RFID/IoT tracking devices, and cloud-integrated fleet monitoring platforms. These systems generate continuous real-time data flows and operate in highly distributed, resource-constrained environments — making them vulnerable to cyber intrusions, manipulation attacks, GPS spoofing, and data-tampering threats.

By applying a multi-stage optimization pipeline that combines model quantization, structured pruning, knowledge distillation, and self-distillation, the original BERT-based intrusion detection model is dramatically compressed while preserving detection performance. The optimized model was comprehensively evaluated on two benchmark datasets: UNSW-NB15 and CIC-IDS2017. Experimental results show that OB-IDS reduces inference time by up to 87.3% and model size/memory footprint by up to 92.6% compared to the full BERT baseline, while maintaining accuracy above 98.1% on both datasets. These findings demonstrate that Transformer-based IDSs can be successfully deployed in real-time threat detection scenarios under severe computational and memory constraints.